How to Turn a Scanned Document into Editable Text (OCR Guide)
A practical OCR guide: how optical character recognition works, how to scan or photograph pages for the best accuracy, and what to do when the result is full of errors.
You have a scanned contract, a photographed page from a book, or a screenshot full of text, and you need the words, not the picture. Retyping is slow and error-prone. OCR, optical character recognition, can do it in seconds, but the quality you get depends heavily on the image you feed it.
How OCR works, in plain words
An OCR engine looks at an image and does three things. First it cleans the image up: converts it to black and white, straightens it and removes noise. Then it finds lines of text and splits them into words and characters. Finally it recognises the characters, using a model trained on huge amounts of printed text and a dictionary for the chosen language to decide between similar-looking options, such as “rn” and “m”.
Modern engines such as Tesseract, which powers our Image to Text tool, use neural networks that read whole lines rather than single letters. On clean printed text they are remarkably accurate, often well above 95 percent of characters.
Why you get errors, and how to avoid them
1. Resolution
OCR needs enough pixels per letter. For normal 10 to 12 point text, scan at 300 DPI. 200 DPI is the practical minimum; 600 DPI rarely helps and makes files huge. For phone photos, fill the frame with the page and avoid zooming in later.
2. Light and shadows
Uneven light is the most common problem with phone photos. Shadows from your hand or phone turn into grey blobs that the engine reads as garbage. Shoot near a window in daylight, hold the phone parallel to the page, and avoid the flash, which creates glare on glossy paper.
3. Skew and curvature
Slightly rotated pages are fine, but strongly tilted photos and the curved pages of a thick book cause line-detection errors. Flatten the page with a sheet of glass or use a phone scanning app that corrects perspective.
4. The wrong language
Choosing the correct language matters more than people think. An English model does not know Cyrillic letters at all and handles German umlauts or French accents poorly. For mixed documents, pick a combination such as English + Russian.
5. Layout
Plain paragraphs are easy. Multi-column magazines, tables and forms are harder: the text may come out in the wrong order. Check the reading order of the result and, for tables, consider re-creating them manually.
6. Handwriting
General-purpose OCR engines are built for printed text. Neat block capitals sometimes work; cursive handwriting usually does not. Handwriting recognition needs specialised models.
Step by step: from paper to editable text
- Scan the pages at 300 DPI, or photograph them in good, even light.
- If you have several photos, combine them into one PDF with JPG to PDF so pages stay in order.
- Run the file through PDF OCR (for PDFs) or Image to Text (for images), choosing the right language.
- Copy the text into your editor and proofread it, paying special attention to numbers, names and amounts.
OCR or PDF to Word?
These solve different problems. If you can already select text in your PDF, it is a digital document and does not need OCR: use PDF to Word to get an editable file with the layout preserved. If you cannot select anything, the PDF is a picture of a page, and OCR is the only way to get the text out.
Privacy matters more than usual here
The documents people OCR are often sensitive: invoices, ID cards, letters from the bank, medical reports. Many online OCR services upload them and keep them for hours or days. Our OCR runs entirely in your browser using WebAssembly. The first run downloads the engine and the language data, a few megabytes, and after that nothing leaves your device.
Our take
OCR quality is decided before you click the button. Ninety percent of bad results we see come from dark, tilted phone photos or the wrong language setting. Take one good photo instead of three quick ones, choose the language, and always proofread numbers. For anything legal or financial, treat OCR output as a draft, never as the final copy.